Z.ai Releases a 321-Billion-Parameter Model as a Free Download

Z.ai has published GLM-5.3-Flash and released the model's weights for download under an MIT license, the permissive terms that allow commercial use and modification.

The model repository on Hugging Face records 321,323,031,390 parameters, split across 62 files, and its configuration is multimodal: it is set up to take images and text as input and return text. The listing gives English and Chinese as the model's languages.

Every benchmark score on the model card is Z.ai's own, and Hugging Face marks all of them unverified. The company reports 84.3 on Terminal-Bench 2.1, 63.4 on DeepSWE and 55.3 on HLE. The card attaches its own conditions to two: the DeepSWE figure came from a run through the mini-swe-agent harness with a 400,000-token context, and the HLE figure was produced with tools switched on across the full set and a 300,000-token context strategy, rather than that benchmark's no-tools default.

TechNode reported on Aug. 27 that Z.ai had identified the model as the one tested anonymously as "Ox Alpha" on the OpenRouter and OpenCode services before release. That identification comes from TechNode, not from the repository, which does not mention it.

TechNode also gives the model 320 billion parameters in total with 18 billion active at any one time; the active figure is the outlet's, and the repository's own metadata does not record it.

Several services are already offering the model for hosted inference, among them Together, Novita and Baseten, according to the repository listing.